Feature extraction

In the feature-extraction configuration, the original classification section is excluded and a new classifier is placed after the retained convolutional and pooling layers.

Programming

From raw data to learnable input

A neural network cannot learn directly from raw sound, images, text, or an unprepared table. Before learning begins, preprocessing changes the data into numerical forms that the network can accept. This preparation includes vectorization, normalization, and decisions about missing values. Inputs and targets are represented as tensors because vectorization produces tensor values, and tensors provide the numerical structure used by the network.

What changes when raw images, text, or other data are converted into tensor values before neural-network processing?

Vectorization means changing a source such as text, an image, or tabular data into numerical values. Those values are organized into tensors before they are passed into the network. The same principle applies to targets: the expected answers must also be represented in a numerical tensor form that the learning process can use.

Why value ranges matter

Neural-network data should usually contain small, relatively homogeneous values. When different features use incompatible ranges, learning can encounter problems. Normalization addresses this by keeping values small and reducing the mismatch between feature ranges.

Preparation	What it does	When to recognize it
Scaling	Maps values into a small range	Useful when the goal is to keep numerical values small
Standardizing	Transforms a feature to a mean of 0 and a standard deviation of 1	Useful when describing each feature relative to its mean and spread

Scaling and standardizing are different ways to prepare numerical features.

What is the difference between mapping values to a small range and transforming them to a mean of 0 with a standard deviation of 1?

Missing values without ambiguity

A dataset may lack a value for a feature in some training or test records. One safe strategy is to represent a missing value as 0, but only when 0 does not already have a meaningful interpretation for that feature. If 0 is a genuine value, using it as the missing marker makes two different situations look identical to the network.

How can missing data be encoded so the network can distinguish an absent value from a genuine value of zero?

Worked example

Choosing a missing-value representation

A feature is absent in some records, but the value 0 is already meaningful for that feature.

1
Check the meaning of zero
Because 0 already represents a real value, it cannot safely serve as the missing marker.
2
Separate the meanings
Choose a representation that lets the network distinguish absence from a genuine zero.
3
Use it consistently
Represent missing entries consistently during training so the network can learn that the marker signals missing data.

Result

Missing data is encoded separately from a meaningful value of zero.

Best practice

Before choosing 0 as a missing marker, ask whether 0 already means something in the feature. If it does, use a representation that preserves the distinction between missing and present.

A pretrained network as a starting point

Definition

A pretrained network is a saved network trained previously on a large dataset, such as ImageNet. Its learned representations can be reused because the network has already learned visual patterns that may also help with a new image problem.

Feature extraction begins before training a classifier for the new image problem. You keep the part of the pretrained network that produces visual representations, pass the new images through that part, and train a separate classifier from scratch on the resulting representations.

How does a new image move from raw pixels through convolutional layers into a reusable feature representation and then to a classifier?

The convolutional base and classifier

The data path has two different jobs. The convolutional base transforms an image into learned visual features. The new classifier interprets those features for the new set of classes. In the feature-extraction configuration, the original classification section is excluded and a new classifier is placed after the retained convolutional and pooling layers.

Which parts provide general features, which part makes the new prediction, and how are they connected?

Depth and feature reusability

The depth of a convolutional layer affects how reusable its features are. Earlier convolutional layers tend to learn local visual patterns such as edges, colors, and textures. Deeper layers tend to represent more abstract patterns, such as a cat ear or a dog eye. Earlier features are generally more reusable, while deeper features may be more tied to the original task.

How do features change from early layers to deeper layers, and why does that affect reuse?

The best reuse boundary depends on how different the new dataset is from the original one. If the datasets differ substantially, the first few convolutional layers may be more appropriate because their features are more general. Higher layers encode increasingly abstract concepts and may be less reusable for a very different task.

Worked example

Selecting a reuse boundary

A new image dataset differs substantially from the dataset used to train the pretrained network.

1
Compare the datasets
A substantial difference suggests that higher-level features may be tied too closely to the original image problem.
2
Prefer general representations
Earlier convolutional layers capture local patterns such as edges, colors, and textures.
3
Place the new classifier
Use the selected convolutional output as the input to a new classifier for the new classes.

Result

Earlier convolutional features may provide a safer reuse boundary when the new dataset is very different.

Common reuse mistakes

Common mistake
Keeping the original classifier for a new set of classes.
What it looks like
The original classifier remains connected even though the new image problem uses different class names.
Why it fails
The original classifier is tied to the classes from the first training task.
Do this instead
Reuse the convolutional base and place a new classifier after its output.
Common mistake
Treating the whole pretrained network as equally reusable.
What it looks like
The deepest layers are selected automatically even when the new dataset differs substantially from the original one.
Why it fails
Deeper layers represent increasingly abstract concepts and may be less reusable for a very different task.
Do this instead
Consider earlier layers when the new dataset is substantially different.
Common mistake
Confusing the convolutional base with the final prediction step.
What it looks like
The visual representation produced by the base is treated as the final answer for the new task.
Why it fails
The base transforms images into features; the new classifier interprets those features for the new classes.
Do this instead
Keep the two jobs separate: feature production in the base and task-specific interpretation in the classifier.
Common mistake
Using zero as a missing marker when zero is meaningful.
What it looks like
A genuine value of 0 and an absent value are both encoded as 0.
Why it fails
The network cannot reliably distinguish a real zero from missing data.
Do this instead
Use a representation that separates missing values from meaningful zeros.

Best practice

A strong reuse decision starts with three questions: Which part produces reusable visual representations? Which part is tied to the original classes? How similar is the new dataset to the original training data?

Guided practice

MEDIUM

A new image dataset has different classes from the original dataset used to train a pretrained network. Describe the complete preparation and feature-extraction path: how the images become tensors, how their values are prepared, how missing values should be considered, which part of the pretrained network should be reused, and where the new classifier belongs.

Hints

Begin with vectorization and tensor representation.
Mention why small, relatively homogeneous values are useful.
Check whether zero has a meaningful interpretation before using it for missing data.
Separate the convolutional base from the original classifier.
Explain how the new classifier uses the base output.

Your attempt (not saved)

What do you think happens?

If the new image dataset differs substantially from the original training dataset, which part is generally the safer starting point for reuse: earlier convolutional layers or deeper convolutional layers?

Earlier convolutional layers
Deeper convolutional layers
The original classifier

Reveal answer

Answer:

Earlier convolutional layers

Earlier layers tend to capture more general local patterns such as edges, colors, and textures. Deeper layers represent more abstract concepts and may be less reusable for a substantially different task.